AI Glossary: 30 Terms for Non-Technical Readers

You’ve read this far—you can use AI and follow the discussions. But AI jargon density is high. This is a 90%-coverage glossary, grouped by theme.

Each term: one-line definition + one-line analogy. After this, you can follow any AI discussion.

I. Foundational Concepts (5)

1. AI (Artificial Intelligence) The umbrella term for making machines exhibit “intelligent” behavior. Analogy: this is an aspirational word, not a specific tech.

2. ML (Machine Learning) Letting machines “figure rules out” from data, instead of humans writing rules. Analogy: show a dog 100 cat photos saying “this is cat”—the dog learns. That’s ML.

3. DL (Deep Learning) A type of ML using multi-layer neural networks. Analogy: ML is the big set; DL is the “heavy-hitter” subset.

4. LLM (Large Language Model) Deep learning model specialized in text, with many parameters (typically 1B+). Analogy: ChatGPT, Claude are LLMs.

5. AGI (Artificial General Intelligence) AI matching or surpassing human ability across most tasks. Doesn’t exist yet. Analogy: current AI is “specialist”; AGI is “generalist.”

II. Model Concepts (6)

6. Parameters The “tunable knobs” inside a model. “7B model” = 7 billion parameters. Analogy: each parameter is a dial; training = adjusting dials.

7. Token The smallest unit a model processes—could be a character, word, or part of a word. Analogy: “Hello” might be 1 token; “你好” might be 2-3 tokens; billing is per-token.

8. Context Window Max tokens the model can “see” at once. Analogy: model’s “short-term memory.” Exceed it, it forgets earlier content.

9. Embedding Converting a word/sentence/document to a vector where semantically similar things are close. Analogy: put each word in a high-dimensional space—“king” and “queen” are near; “king” and “banana” are far.

10. Transformer The 2017 neural network architecture; the foundation of all modern large models. Analogy: the engine. ChatGPT, Claude, Gemini are different “cars” sharing the Transformer engine.

11. Attention (Mechanism) Transformer’s core: each token “looks back” at others, weighing by relevance. Analogy: reading “the cat chased the ball because it was curious”—your eye darts back to “cat” to figure out “it.” That’s attention.

III. Training Pipeline (5)

12. Pre-training Training a base model on massive internet text (hundreds of billions of tokens). Analogy: making a child read the entire library. GPT-3’s pretraining cost ~$4.6M.

13. Fine-tuning Continuing training on top of a pretrained model, with domain-specific data. Analogy: the same child reads medical textbooks for a week, becoming a medical assistant.

14. SFT (Supervised Fine-Tuning) Using “question-answer” pairs to teach the model to follow instructions. Analogy: show the child 1,000 “good question-answer examples,” let them learn “how should I respond.”

15. RLHF (Reinforcement Learning from Human Feedback) Have human raters score multiple AI answers; the model learns “what humans prefer.” Analogy: ChatGPT’s “polite, warm, detailed” responses come from RLHF.

16. LoRA (Low-Rank Adaptation) A lightweight fine-tuning method—only updates a small subset of parameters, saving 90% compute. Analogy: don’t rebuild the body; just add patches at the joints.

IV. Inference & Use (6)

17. Inference The model generating an answer (vs training). Analogy: training = “learning”; inference = “exam.”

18. Prompt Your instruction to the AI. Analogy: like briefing a new employee—clear briefing = good work.

19. Prompt Engineering Crafting prompts to get more precise AI output. Analogy: the art of efficiently “directing” AI. See L0-05.

20. CoT (Chain of Thought) Have the AI think step-by-step before answering. Analogy: tell it “reason step by step”—its accuracy immediately improves.

21. Few-shot Include several examples in the prompt for AI to imitate. Analogy: “write in the style of these 3 examples”—more effective than “be lively.”

22. Hallucination AI confidently saying wrong things, making things up. Analogy: an amnesiac storyteller—it thinks it remembers, but it’s guessing.

V. Augmentation & Extension (5)

23. RAG (Retrieval-Augmented Generation) First retrieve relevant docs from a knowledge base, then let the AI answer based on the docs. Analogy: at exam time, instead of testing from memory, give the AI a textbook to flip through—accuracy soars.

24. Agent An AI system that can plan, call tools, and act over multiple steps autonomously. Analogy: from “an assistant that answers one question at a time” to “an intern who can schedule meetings, book flights, write emails.”

25. Tool Use AI calling external tools mid-conversation—calculator, web search, APIs. Analogy: can’t compute exactly? Let AI open the “calculator app.”

26. MCP (Model Context Protocol) Anthropic’s standard protocol for AI to safely connect to external tools and data. Analogy: think of it as “AI world’s USB port.”

27. Multimodal AI handling multiple input types simultaneously—text, image, audio, video. Analogy: AI that can “see, hear, and read”—no longer just reading words.

VI. Performance & Systems (4)

28. Quantization Reducing parameter precision (e.g., 32-bit float → 4-bit integer)—smaller model, faster inference. Analogy: compressing a HD photo to JPEG—slight quality loss, much less space.

29. KV Cache Model caching intermediate results during generation, avoiding recomputation. Long conversations get slower because the cache grows. Analogy: keeping your past notes spread on the table during a presentation, instead of mentally recalling everything.

30. Open-source vs Closed

• Open: model weights public; anyone can download, modify, deploy locally. (Llama, Mistral, DeepSeek, Qwen)
• Closed: only via API. (GPT-4, Claude, Gemini) Analogy: open-source is Linux (free but DIY); closed is Windows (polished but locked).

A “See It, Understand It” Cheat Sheet

The next time you see these combinations:

Jargon	Translation
”This model has 70B parameters"	"Its brain has 70 billion knobs"
"200k context"	"It can read about 150K Chinese characters at once"
"We use RAG"	"We gave the AI a knowledge base"
"It hallucinates"	"It makes things up confidently"
"LoRA fine-tuning"	"Light training on a small dataset"
"Tool use integrated"	"AI can now call other software itself"
"Token cost $0.003"	"A few cents per sentence"
"MoE architecture"	"Multiple ‘experts’ inside, only some active each call"
"Inference acceleration"	"Make AI answer faster"
"Embedding-based retrieval"	"Search docs by semantic similarity”

This Is Just the Start

L1-L7 paths use lots of these terms. Bookmark this—come back when stuck.

Next: “L0 Graduation: What’s Next?”—we look back, and look forward.